[Build] Align CPython 3.14 CUDA 13 wheels with TensorRT 11

[Tenshock]

Describe the issue

The official onnxruntime-gpu==1.29.0 CPython 3.14 CUDA 13 wheels that include TensorrtExecutionProvider still link TensorRT 10 on both supported x64 platforms:

• Linux x86_64 requests libnvinfer.so.10 and libnvonnxparser.so.10.
• Windows x64 requests nvinfer_10.dll and nvonnxparser_10.dll.

They therefore cannot load TensorrtExecutionProvider in a CPython 3.14 environment using the current TensorRT 11 CUDA 13 distribution, which supplies the TensorRT 11 ABI.

I propose moving the TensorRT-enabled CUDA 13 Python-wheel profile to CUDA 13.3 and TensorRT 11.2.1.2 coherently:

• Use CUDA 13.3 and cuDNN 9 for Windows x64 and Linux x86_64 wheels.
• Build the Windows x64 and Linux x86_64 TensorRT providers against TensorRT 11.2.1.2.
• Apply the version change to every Python ABI built by the CUDA 13 packaging pipeline, including CPython 3.14.
• Leave Linux aarch64, CUDA 12, and non-Python package pipelines unchanged. The current Linux aarch64 wheel does not contain a TensorRT provider, so it is not a target for this ABI fix.

NVIDIA publishes TensorRT 11.2.1.2 packages built against CUDA 13.3 for Linux x86_64 and Windows x64.

Urgency

This blocks TensorrtExecutionProvider in CPython 3.14 environments using the current TensorRT 11 CUDA 13 distribution.

Target platforms

• Linux x86_64: CPython 3.14 (cp314-cp314 and cp314-cp314t)
• Windows x64: CPython 3.14 (cp314-cp314)
• CUDA 13.3, cuDNN 9, and TensorRT 11.2.1.2 on both platforms

Build scripts

The official pipeline path is:

• tools/ci_build/github/azure-pipelines/py-cuda13-packaging-pipeline.yml
• tools/ci_build/github/azure-pipelines/stages/py-gpu-packaging-stage.yml
• tools/ci_build/github/azure-pipelines/stages/py-linux-gpu-stage.yml
• tools/ci_build/github/azure-pipelines/stages/py-win-gpu-stage.yml
• tools/ci_build/github/azure-pipelines/templates/jobs/download_win_gpu_library.yml
• tools/ci_build/github/linux/build_linux_python_package.sh

The current pipeline uses CUDA 13.0. Its x64 TensorRT inputs are TensorRT 10.x.

The CUDA 13 packaging pipeline therefore needs a CUDA 13.3 Linux x86_64 build image and matching Windows CUDA, cuDNN, and TensorRT inputs.

Error / output

Linux wheel:

onnxruntime_gpu-1.29.0-cp314-cp314-manylinux_2_28_x86_64.whl

SHA-256:

5c50d24541bf61e668dfe0a7a3a62de98def81959c832cc2d376814b661cb006

Relevant readelf -d libonnxruntime_providers_tensorrt.so output:

Shared library: [libcublas.so.13]
Shared library: [libnvinfer.so.10]
Shared library: [libnvonnxparser.so.10]
Shared library: [libcudart.so.13]
Shared library: [libcudnn.so.9]

Windows wheel:

onnxruntime_gpu-1.29.0-cp314-cp314-win_amd64.whl

SHA-256:

4247025e674a666f839c02b90c829277cb2e1be11576d3646cad5c86b7fc26c5

Relevant objdump -p onnxruntime_providers_tensorrt.dll output:

DLL Name: cublas64_13.dll
DLL Name: nvinfer_10.dll
DLL Name: nvonnxparser_10.dll
DLL Name: cudnn64_9.dll

Expected TensorRT dependencies:

Linux:   libnvinfer.so.11, libnvonnxparser.so.11
Windows: nvinfer_11.dll, nvonnxparser_11.dll

The CPython 3.14 Linux aarch64 wheel was also inspected and contains CUDA and shared providers, but no TensorRT provider; it is therefore outside this issue's target platform set.

GCC / Compiler Version

The current Linux x86_64 CUDA 13 packaging image uses GCC 14.

References

• https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/release-notes-11/11.2.1.html
• https://docs.nvidia.com/deeplearning/tensorrt/latest/installing-tensorrt/prerequisites.html
• https://github.com/NVIDIA/TensorRT/blob/main/README.md
• Nightly packages for Python 3.14 or CUDA 13.x available for testing #26547
• Guard removed APIs in TRT 11 #28611
• https://pypi.org/project/onnxruntime-gpu/1.29.0/

[Tenshock]

Implementation drafts:

• [Build] Parameterize CUDA and TensorRT Python wheel inputs #32279 parameterizes the CUDA and TensorRT Python-wheel inputs.
• [TensorRT] Build CUDA 13 Python wheels with TensorRT 11 #32280 verifies the TensorRT provider ABI dependencies and runs TensorRT inference with CPU fallback disabled.

The CUDA 13 version-selection commit needs the Linux x86_64 CUDA 13.3 GCC 14 build-image tag.

Please also confirm that these Windows build assets are available to the Python packaging pipeline:

• Windows x64 CUDA 13.3 SDK
• Windows x64 TensorRT 11.2.1.2 CUDA 13.3 package